-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathget_saliency.py
executable file
·150 lines (137 loc) · 7.26 KB
/
get_saliency.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
import os
import argparse
import numpy as np
import timm
import cv2
import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision.utils import save_image
import torchvision.datasets as datasets
from torch.utils.data import DataLoader
from submodule.pytorch_smoothgrad.lib.gradients import SmoothGrad
from submodule.pytorch_smoothgrad.lib.image_utils import save_diff_map, save_as_heatmap, save_as_overlay
from src.utils import binary_accuracy, CosineAnnealingWithWarmup, get_args_parser, get_transform_wo_crop
from PIL import Image
from src.models import LinearModel, LinearProbeModel
from run_linear_probe import extract_features, evaluate_linear_probe
from src.perspective_data import PerspectiveDataset, FeaturesDataset
from tqdm import tqdm
def parse_args():
parser = get_args_parser()
parser.add_argument('--img_dir', type=str, default='',
help='Input image path')
parser.add_argument('--out_dir', type=str, default='./output/',
help='Result directory path')
parser.add_argument('--n_samples', type=int, default=10,
help='Sample size of SmoothGrad')
parser.add_argument("--ckpt_path", type=str, help="Model Checkpoint Path")
parser.add_argument("--probe_ckpt", type=str)
args = parser.parse_args()
return args
def get_finetune_saliency(img_name, device, args):
img_name = '_'.join(args.img_dir.split('/')[-3:]).split('.')[0]
if args.ckpt_path == None:
model = timm.create_model(args.model_name, pretrained=True, num_classes=1)
else:
model = timm.create_model(args.model_name, pretrained=False, num_classes=1)
model.load_state_dict(torch.load(args.ckpt_path).state_dict())
#model.eval()
data_config = timm.data.resolve_model_data_config(model)
transform = get_transform_wo_crop(data_config)
img = Image.open(args.img_dir).convert('RGB')
preprocessed_img = torch.unsqueeze(transform(img), 0)
model = model.to(device)
smooth_grad = SmoothGrad(
pretrained_model=model,
cuda=True,
n_samples=args.n_samples,
magnitude=True)
smooth_saliency = smooth_grad(preprocessed_img, index=None)
#save_as_gray_image(smooth_saliency, os.path.join(args.out_dir, 'smooth_grad.jpg'))
save_as_heatmap(smooth_saliency, os.path.join(args.out_dir, f'{args.model_name}_ft_{img_name}_grad.png'))
img = cv2.resize(cv2.imread(args.img_dir, 1), smooth_saliency.shape[1:])
save_as_overlay(img, smooth_saliency, os.path.join(args.out_dir, f'{args.model_name}_ft_{img_name}.png'))
return smooth_saliency
def train_linear_probe(model, train_loader, val_loader, criterion, optimizer, lr_scheduler, device, args):
best_acc_val = 0
best_acc_train = 0
for epoch in tqdm(range(args.epochs)):
model.train()
epoch_acc = []
epoch_loss = []
for i, batch in enumerate(train_loader):
features, labels = batch
features = features.to(device)
labels = labels.float().to(device)
labels = torch.unsqueeze(labels, 1)
optimizer.zero_grad()
preds = model(features)
loss = criterion(preds, labels)
loss.backward()
optimizer.step()
acc = binary_accuracy(preds, labels)
epoch_acc.append(acc)
epoch_loss.append(loss.item())
with torch.no_grad():
val_acc, val_loss = evaluate_linear_probe(model, val_loader, criterion, device, False, args)
train_acc = sum(epoch_acc)/float(len(epoch_acc))
if val_acc > best_acc_val:
torch.save(model, f'./logs/linear_probe_ckpts/{args.model_name}_{args.task}.ckpt')
best_acc_val = val_acc
best_acc_train = train_acc
return best_acc_train, best_acc_val
def get_linear_probe_saliency(img_name, device, args):
model = timm.create_model(args.model_name, pretrained=True, num_classes=0)
model = model.to(device)
model.eval()
data_config = timm.data.resolve_model_data_config(model)
transform = get_transform_wo_crop(data_config)
if args.probe_ckpt is None:
train_dataset = datasets.ImageFolder(os.path.join(args.data_dir, 'train_flip'), transform=transform)
val_dataset = datasets.ImageFolder(os.path.join(args.data_dir, 'val'), transform=transform)
train_loader = DataLoader(train_dataset, batch_size=args.extract_batch_size, num_workers=args.num_workers, pin_memory=True, drop_last=False)
val_loader = DataLoader(val_dataset, batch_size=args.extract_batch_size, num_workers=args.num_workers, pin_memory=True)
train_features, train_labels = extract_features(model, train_loader, device)
val_features, val_labels = extract_features(model, val_loader, device)
train_feat_dataset = FeaturesDataset(train_features, train_labels)
val_feat_dataset = FeaturesDataset(val_features, val_labels)
train_feat_loader = DataLoader(train_feat_dataset, batch_size=args.batch_size,
num_workers=args.num_workers, shuffle=True, drop_last=True)
val_feat_loader = DataLoader(val_feat_dataset, batch_size=args.batch_size, num_workers=args.num_workers)
linear_model = LinearModel(train_features.shape[-1], args.num_classes, args.dropout_rate)
linear_model = linear_model.to(device)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.AdamW(linear_model.parameters(), lr=args.learning_rate,
weight_decay=args.weight_decay, amsgrad=False)
best_acc_train, best_acc_val = train_linear_probe(linear_model, train_feat_loader, val_feat_loader, criterion, optimizer, None, device, args)
print("Best acc validation", best_acc_val)
print("Best acc train", best_acc_train)
else:
linear_model = torch.load(args.probe_ckpt)
#linear_model.load_state_dict(torch.load(args.probe_ckpt).state_dict())
model.train()
full_model = LinearProbeModel(model, linear_model)
img = Image.open(args.img_dir).convert('RGB')
preprocessed_img = torch.unsqueeze(transform(img), 0)
smooth_grad = SmoothGrad(
pretrained_model=full_model,
cuda=True,
n_samples=args.n_samples,
magnitude=True)
smooth_saliency = smooth_grad(preprocessed_img, index=None)
#save_as_gray_image(smooth_saliency, os.path.join(args.out_dir, 'smooth_grad.jpg'))
save_as_heatmap(smooth_saliency, os.path.join(args.out_dir, f'{args.model_name}_lp_{img_name}_grad.png'))
img = cv2.resize(cv2.imread(args.img_dir, 1), smooth_saliency.shape[1:])
save_as_overlay(img, smooth_saliency, os.path.join(args.out_dir, f'{args.model_name}_lp_{img_name}.png'))
return smooth_saliency
def main():
args = parse_args()
device = torch.device(f'cuda:0')
img_name = '_'.join(args.img_dir.split('/')[-3:]).split('.')[0]
ft_map = get_finetune_saliency(img_name, device, args)
lp_map = get_linear_probe_saliency(img_name, device, args)
img = cv2.resize(cv2.imread(args.img_dir, 1), ft_map.shape[1:])
save_diff_map(ft_map, lp_map, img, os.path.join(args.out_dir, f'{args.model_name}_diff_{img_name}_grad.png'),os.path.join(args.out_dir, f'{args.model_name}_diff_{img_name}.png'))
if __name__ == "__main__":
main()